Resources & Reproducibility

The CIMILab Reproducibility Standard

Every project graduating from CIMILab to a public release or publication must satisfy the following pipeline requirements. This ensures our baselines are strong and our results can be verified by independent researchers.

• Versioned Data All training, validation, and test splits are versioned with checksums (e.g., using DVC). Data preprocessing scripts are deterministic and included in the repo.
• Fixed Random Seeds Random seeds for data shuffling, model initialization, and augmentation are explicitly set and documented for identical reruns.
• Config-Driven Experiments Hyperparameters are stored in configuration files (YAML/JSON) rather than hardcoded, and the exact config used for the published result is committed to the repository.
• Scripted Evaluation A single command (e.g., bash eval.sh) can run the trained model on the test set and reproduce the exact tables/metrics reported in the paper.
• Released Weights Unless restricted by data-privacy agreements (common in medical AI), pre-trained model weights are released on HuggingFace or Zenodo.